Deep spatial-constraints networks for unsupervised anomaly detection in multivariate time series data

High-dimensional time series anomaly detection has always been an important challenge in the field of system security. Most existing methods are dedicated to modelling the temporal variation of features to capture anomalous moment points, however as features become more high-dimensional, the associations between features take on a complex spatial structure. This spatial structure information will compensate for the constraints of unsupervised training conditions, and guide the model to be more fully trained. In this study, we propose a detection model that incorporates spatial supervision signals. The model not only simultaneously models the temporal and spatial dependencies, but also simulates the topological structure and physical characteristics of data in the real world through graph structure learning and contrastive learning, providing guidance for anomaly detection. We conducted experiments on two real-world datasets and demonstrated that our model outperforms the baseline. Finally, we conducted detailed data analysis to provide interpretability for the model.